Upload
others
View
0
Download
0
Embed Size (px)
Citation preview
Parametric Robust H2 Control Design with
Generalized Multipliers via LMI Synthesis
David Banjerdpongchai�
Durand Bldg., Room 110
Dept. of Electrical Engineering
Email: [email protected]
Jonathan P. How
Durand Bldg., Room 277
Dept. of Aeronautics and Astronautics
Email: [email protected]
Stanford University, Stanford CA 94305
August 1, 1997
Abstract
This paper presents a new, combined analysis and synthesis procedure that providesa less conservative robust control design technique for systems with real parametric un-certainty. The robust stability for these systems is analyzed by the passivity theoremwith generalized multipliers, and the worst case H2 performance is investigated usingan upper bound on the total output energy. The dynamics of the multipliers are sys-tematically chosen using knowledge from the linear part of the uncertain systems. Thisapproach provides additional degrees of freedom in the synthesis that lead to a reduc-tion of the conservatism in the worst case H2 performance and achieved robustnessbounds. However, the formulation of the control design problem is very complicatedand it is di�cult to solve directly. This paper presents an iterative algorithm, whichis an H2 equivalent of the D{K iteration for the �=Km synthesis, to account for thecomplicated couplings in the synthesis problem. We use a simple beam system with anuncertain modal frequency to illustrate that this synthesis technique with generalizedmultipliers results in less conservative controllers than previously published Popov con-troller synthesis techniques. In the process, we demonstrate that this design approachis very e�ective and simple to implement numerically.
Keywords: real parametric uncertainty; passivity theorem; plant{dependent multi-plier; controller synthesis; bilinear matrix inequality; linear matrix inequality.
�Author to whom all correspondence should be addressed. Tel: (650) 723-9833; Fax: (650) 723-8473.
1
1 Introduction
Reducing conservatism in the robust stability analysis and synthesis for systems with real
parametric uncertainty is currently a key issue in the controls community, but its heritage
dates back to the 1960's. The primary focus of the early work [Zames, 1966a, Zames, 1966b,
Sandberg, 1964] was not directly on the issue of the parametric uncertainty, but more on
developing a fundamental understanding of the multivariable stability analysis based on the
conic-sector, positivity, and loop gain concepts. This led more recently to quantitative mea-
sures such as the multivariable stability marginKm [Safonov, 1982], and the structured singu-
lar value � [Doyle, 1982]. This analysis was extended to the robust controller synthesis using
variants of the so-called D{K iteration [Safonov, 1983, Doyle, 1983], which have recently
been improved by including real parametric uncertainty [Fan et al., 1991, Young, 1993] and
by eliminating the need for curve-�tting [Safonov and Chiang, 1993]. The �=Km synthesis
problem using Bilinear Matrix Inequalities (BMIs) is formulated in [Safonov et al., 1994,
Goh et al., 1994]. The BMI approach has been shown to improve the guaranteed lower
bounds of the multivariable stability margins by 10% over the corresponding results from
the D{K iteration, with no increase in the controller order. A key advantage of the BMI
technique is that it enables control engineers to address several other open problems of the
robust control synthesis, namely, �xed order control synthesis, and decentralized controller
architecture.
Balakrishnan presented a uni�ed framework for robust stability tests based on the pas-
sivity theorem with multipliers [Balakrishnan, 1995], and also demonstrated that these tests
can be performed using convex optimization over Linear Matrix Inequalities (LMIs). Recent
works [Boyd et al., 1994, Feron, 1994, How, 1993] on the robust H2 performance analysis
for uncertain systems have also focused on including these stability multipliers. While the
actual worst case H2 performance is very di�cult to compute, its upper bound is much easier
to calculate. These authors have demonstrated that the conservatism of the bound for the
worst case H2 performance could be reduced by applying the passivity theorem with gen-
eralized multipliers, and that the formulation of the analysis tests naturally lead to LMIs.
These robust analysis approaches have been tested for simple systems using generalized,
dynamic multipliers. However, they have not been extended to the design of robust con-
trollers that guarantee the robust stability andH2 performance. Although the robust designs
with Popov multipliers (or a slight generalization of Popov multipliers for repeated uncer-
tainty) has been previously shown by several authors, such as [Haddad and Bernstein, 1991,
How, 1993, How et al., 1994, Sparks and Bernstein, 1995, Banjerdpongchai and How, 1996],
these controllers are designed for systems with sector bounded nonlinear uncertainty, which
could be a source of conservatism when working with systems for real parametric uncertainty.
In this paper, we introduce a new, combined analysis and synthesis procedure that leads
to an e�ective robust control design technique and demonstrate that this technique can be
used to design less conservative controllers for systems with real parametric uncertainty. Our
2
approach is motivated by the work of [Haddad et al., 1992, Balakrishnan, 1995, Feron, 1994].
The direct extension of the robust analysis to the controller synthesis results in BMIs, which
currently are di�cult to solve directly [Toker and �Ozbay, 1995]. As a result, El Ghaoui and
Balakrishnan proposed an iterative procedure for solving BMI problems using a two-stage
optimization process, called the V{K iteration [El Ghaoui and Balakrishnan, 1994]. Some
of the design variables in the BMIs are �xed during each phase of this iteration, leading to
LMIs in the remaining variables. This technique has been shown to work well on simple
examples, but on complicated objectives, such as robust control designs to minimize an
H2 performance, this approach has been found to converge very slowly, if at all. This
synthesis algorithm was recently improved leading to a systematic control design for systems
with unstructured uncertainty [El Ghaoui and Folcher, 1996]. We have already successfully
applied an extension of this algorithm to the parametric robustH2 control design with Popov
multipliers [Banjerdpongchai and How, 1996]. Key advantages of the LMI formulation when
compared to the previously used gradient techniques are the simplicity and low overhead
in the numerical implementation. In this paper, we extend the synthesis to the case of
generalized multipliers, which is an important step for developing less conservative controllers
for systems with real parametric uncertainty.
Our design objective is achieved by combining the passivity analysis using multipliers
and the worst case H2 performance, i.e., the output energy, of an LTI system subject to
real parametric uncertainty. In the process we show the di�culties that arise when we
simultaneously select the optimal parameters for both the multiplier and the compensator.
We take advantage of the closed-loop matrix structure to eliminate some design parameters
from the problem formulation using a simple algebraic technique. Although the problem
size and the number of design parameters are reduced, some couplings still remain. Hence,
we apply an iterative algorithm using LMI synthesis tools [Vandenberghe and Boyd, 1994,
Wu and Boyd, 1996] to solve the design problem. As we will show, this approach is quite
distinct from the D{K iteration for the �=Km synthesis because some variables are shared
between the two main stages of the iterative solution.
The paper is organized as follows. In the next section, we present the various mathe-
matical notations and lemmas used in this paper. The robust controller synthesis is quite
complex and the algorithm draws on a variety of techniques to address the multiplier se-
lection as well as the controller parameterization. The synthesis formulation using these
previously published techniques is developed in x3. The solution procedure and algorithm
are presented in x4. Lastly, we use the Bernoulli Euler Beam system to demonstrate that
the synthesis with generalized multipliers leads to less conservative controllers for systems
with real parametric uncertainty.
3
2 Preliminaries
The following brie y summarizes the key notations which we will use to present the main
theoretical results in x3.1 and x3.2.
R denotes the set of real numbers. R+ denotes the set of nonnegative real numbers.
Rm�n is the vector space of m � n real matrices. For any matrix A 2 Rm�n, AT denotes
its transpose and A� denotes its complex conjugate transpose. A? denotes an orthogonal
complement of A, i.e., ATA? = 0 and [A A?] is of maximum rank. The identity matrix is
denoted by I. If A 2 Rn�n, TrA denotes the trace of A. If A is square and invertible, then
A�1 is its inverse. Given a set of matrices A1 2 Rn1�n1 ; : : : ; AN 2 RnN�nN , and n =PN
i=1 ni,
diag(A1; : : : ; AN) denotes the n� n matrix
26666664
A1 0 : : : 0
0 A2. . .
......
. . . . . . 0
0 : : : 0 AN
37777775:
When there is no ambiguity, diagi(Ai) or diagNi=1(Ai) denotes diag(A1; : : : ; AN). For any
two matrices A and B 2 Rn�n, the inequality A < B (A � B) means that both A and B
are symmetric and that B � A is positive de�nite (positive semide�nite).
Ln2 is the Hilbert space of square-integrable signals de�ned over R+ with n components,
i.e., u 2 Ln2 satisfying
R1
0 uTu dt < 1. Ln2 is often abbreviated as L2. A causal n-input
n-output operator F : Rn ! Rn is said to be L2 stable if there exist � 0 and � such that
kFuk2 � kuk2 + �; 8 u 2 L2; (1)
where k � k2 is de�ned as the L2 norm. The L2 gain of F is de�ned as the smallest such
that (1) holds for some �. For the linear operator F , we have the following de�nitions. I is
the identity operator; F�1 is the inverse of F , i.e., FF�1 = I; F � is the adjoint of F ; and
F�� is the inverse of the adjoint of F . F is said to be passive if
Z T
0u(t)T (Fu)(t) dt � 0; 8T � 0; 8 u 2 L2:
It is strictly passive if it satis�es
Z T
0u(t)T (Fu)(t) dt > 0; 8T � 0; 8 u 2 L2:
Let F be an LTI system with a transfer function F (s). Suppose F (s) is stable, i.e., all poles
are on the open left half of s-plane. In the frequency domain, the condition for F (s) to be
4
passive is that
F (j!) + F (j!)� � 0; 8! 2 R: (2)
F (s) is said to be positive real if F (s) satis�es (2). F (s) is strictly positive real if it satis�es
F (j!) + F (j!)� > 0; 8! 2 R:
The state-space condition for a passive LTI system is given in the following lemma. The
lemma [Anderson and Vongpanitlerd, 1973, Chapter 5{7] is previously stated as in matrix
equations, but in this paper we will use the form of a linear matrix inequality.
Lemma 2.1 (Positive Real Lemma [Boyd et al., 1994, pages 34{35]) Let F (s) be a transfer
matrix of a stable LTI system, with the minimal realization fA;B;C;Dg. F (s) is positive
real or passive if and only if there exist P = P T > 0 satisfying
"PA+ ATP PB � CT
BTP � C �(D +DT )
#� 0:
Proof: See [Boyd et al., 1994, pages 34{35]. 2
We note that F (s) is strictly positive real or strictly passive if there exists P = P T > 0
satisfying "PA+ ATP PB � CT
BTP � C �(D +DT )
#< 0:
The following lemmas will be very useful in developing the controller design technique in x4.
Lemma 2.2 (Elimination Lemma [Boyd et al., 1994, page 32]) Let G 2 Rn�n, U 2 Rn�p
and V 2 Rn�q. There exists a matrix X 2 Rp�q such that
G+ V XTUT + UXV T < 0;
if and only if
V T?GV? < 0; UT
?GU? < 0:
Proof: See [Boyd et al., 1994, pages 32{33]. 2
Lemma 2.3 (Completion Lemma [Packard et al., 1991]) Let P and Q 2 Rm�m be positive
de�nite matrices. There exists a positive de�nite matrix ~P 2 R2m�2m such that the upper
left m�m block of ~P is P , and that of ~P�1 is Q, if and only if
"P I
I Q
#� 0: (3)
Proof: See [Packard et al., 1991]. 2
5
For each pair of matrices P and Q that strictly satisfy (3), the set of matrices ~P satisfying
the conditions in Lemma 2.3 is parameterized by
~P =
"I 0
0 MT
# "P I
I (P �Q�1)�1
# "I 0
0 M
#;
where M 2 Rm�m is an arbitrary invertible matrix. Then,
~Q = ~P�1 =
"I 0
0 NT
# "Q I
I (Q� P�1)�1
# "I 0
0 N
#;
where N = (I �QP )M�1.
3 Problem Statement
We consider an LTI system, i.e., the nominal system G, subject to the uncertainty � (see
Figure 1, where K is not considered) described by
_x = Ax+Bpp+Bww +Buu
q = Cqx +Dqpp+Dqww +Dquu
z = Czx +Dzpp+Dzww +Dzuu
y = Cyx+Dypp+Dyww +Dyuu
p = ��q
(4)
where x : R+ ! Rn is the state, u : R+ ! Rnu is the control input, w : R+ ! Rnw is the
disturbance input, y : R+ ! Rny is the measured output, z : R+ ! Rnz is the performance
output, q : R+ ! Rnp and p : R+ ! Rnp are the input/output of the uncertainty �.
The uncertainty � is assumed to be a diagonal constant matrix with positive elements, i.e.,
� 2 �, where
� :=n� : � = diag(�1; : : : ; �np); and �i > 0; 8 i = 1; : : : ; np:
o:
In control theory, this is referred to as real parametric uncertainty.
For well-posedness, Dzw is assumed to be identically zero. To signi�cantly simplify the
analysis and synthesis, we assume Dzp, Dqw, and Dqu are identically zero.
Remark 3.1 We note that this formulation can be easily extended to handle the system
with the constant diagonal uncertainty � with elements satisfy j�ij < ; 8 i = 1; : : : ; np. In
these cases, we apply a bilinear sector transformation [Desoer and Vidyasagar, 1975, pages
50{52] so that the stability of the uncertain system can be analyzed by the passivity theorem.
In such cases, we assume that ( I + �) is invertible [Desoer and Vidyasagar, 1975, pages
6
K
Gy
q
u
�
w zp{1
Figure 1: Elements of the robust synthesis problem
219{224] and de�ne Gqp(s) to be the transfer matrix from p to q. After the transformation,
the uncertain system (4) is described by
~Gqp(s) = (I � Gqp(s))�1(I + Gqp(s));
~� = ( I ��) � ( I +�)�1;
where \�" denotes a composition operator. As discussed in [Anderson, 1972], it can be shown
that � has an L2 gain less than (i.e., j�ij < ) if and only if ~� is strictly passive (i.e.,~�i > 0). Given Gqp with a state-space realization fA;Bp; Cq; Dqpg, the state-space realization
of ~Gqp is
fA+ Bp(I � Dqp)�1; 2Bp(I � Dqp)
�1; (I � Dqp)�1 Cq; (I + Dqp)(I � Dqp)
�1g:
Other classes of the uncertainty, such as a diagonal passive operator or a diagonal passive
LTI uncertainty, can be handled in a similar manner [Balakrishnan, 1995]. An extension to
the classes of uncertainty with elements having an L2 gain less than is also straightforward
via the bilinear sector transformation.
The objective of this paper is to design a strictly proper full order LTI controller using
multiplier theory for the uncertain system (4) such that the robust stability of the system
is achieved and an upper bound of the worst case H2 performance is minimized. The for-
mulation is quite complicated because it requires a simultaneous selection of the optimal
parameters of both the multipliers and compensators. In the following subsections, we will
develop the robust H2 synthesis formulation from the robust H2 analysis with multipliers.
3.1 Absolute Stability Analysis with Multipliers
The robust stability analysis is based on the passivity theorem with multipliers, i.e., multi-
plier theory [Zames, 1966a, Zames, 1966b, Desoer and Vidyasagar, 1975]. These multipliers
7
are devised to capture additional information, i.e., structure and type, of the uncertainty �
in order to obtain less conservative conditions for robustness analysis. The application of
multiplier theory is given in the following theorem. The bene�ts of including this additional
information will be discussed in the numerical example x5.
Theorem 3.1 ([Desoer and Vidyasagar, 1975, pages 203]) Consider the system (4) and as-
sume it has a solution x 2 L2. Let u = 0, and w = 0. Suppose there exists an operator
W : L2 ! L2 satisfying the following conditions.
1. W can be factored such that W = W�W+, where W�, W+, and their inverses map L2
to L2;
2. W� is linear, hence W �� and W��
� are well-de�ned;
3. W , W+, W��, and their inverses have �nite gains; and
4. W+, W��, and their inverses are causal.
If WGqp has a �nite gain, WGqp is strictly passive, and �W�1 is passive, then the uncertain
system (4) is L2 stable.
Proof: See [Desoer and Vidyasagar, 1975, pages 203{204]. 2
W is called the stability multiplier. Because W� and its inverse are anticausal; and W+
and its inverse are causal [Desoer and Vidyasagar, 1975, Balakrishnan, 1995], the multipli-
ers satisfying the conditions in Theorem 3.1 are noncausal. Furthermore, because � and Gqp
are causal and W is noncausal, WGqp and �W�1 are noncausal. [Feron, 1994] discusses a
su�cient condition with many theoretical steps to search for a noncausal multiplier W such
that WGqp is strictly passive and the performance bound is achieved. This approach results
in a sophisticated optimization problem over LMIs. However, [Desoer and Vidyasagar, 1975]
show that for the multipliers satisfying the conditions in Theorem 3.1, WGqp (i.e., a non-
causal operator) is strictly passive if and only ifW+GqpW��� (i.e., a causal operator) is strictly
passive. Moreover, �W�1 is passive if and only ifW ���W
�1+ is passive. [Balakrishnan, 1997]
considers the robust analysis of the general framework, i.e., the operator W+GqpW��� , and
formulates the test as a convex optimization over LMI constraints. The underlying numer-
ical methods are based on a state-space approach and result in guaranteed performance
bounds. This general framework might o�er an alternative way to design robust controllers.
However, the conservatism of these bounds obtained by this general framework remains
for further investigation. For simplicity in the following synthesis formulation, we set W�
equal to identity, so that W = W+ (i.e., the set of causal multipliers). Although this
choice of multipliers is more restricted than the generalized (noncausal) multipliers, this
choice allows far more freedom than the Popov multipliers that have been investigated pre-
viously [Banjerdpongchai and How, 1996]. Future research will explore the advantages of
using the factorized form of the generalized multipliers in the robust analysis/synthesis.
8
In practice, a �nite dimensional approximation of the set of multipliers is used to test the
assumptions in Theorem 3.1 [Safonov and Chiang, 1993, Balakrishnan, 1995, Feron, 1994,
Balakrishnan, 1997]. To capture real parametric uncertainty � 2 �, we select the multiplier
W which has the form
W =
8>><>>:diag(W1; : : : ;Wnp);
where Wi is linear time-invariant, �nite-dimensional,
stable, positive real and has no poles on the imaginary axis.
(5)
Note that the freedom in selecting this multiplier serves as a qualitative measure of the
conservatism of the robustness test. The multipliers of the form (5) include very general
parameterizations of the stability multipliers involving RL (resistor{inductor), RC (resistor{
capacitor), and shifted LC (inductor{capacitor) [How and Haddad, 1994]. We denote the
state-space realization of Wi as fAWi; BWi
; CWi; DWi
g, then the multiplier W is described by
_xW = AWxW +BW q;
qW = CWxW +DW q;(6)
where xW : R+ ! RnW is the multiplier state, xW;0 = 0 and
AW = diagnpi=1(AWi
); BW = diagnpi=1(BWi
);
CW = diagnpi=1(CWi
); DW = diagnpi=1(DWi
):
Remark 3.2 Other types of the uncertainty can be handled by choosing appropriate classes
of multipliers [Balakrishnan, 1995]. For example, a constant diagonal positive de�nite ma-
trix is chosen for the uncertainty that is a diagonal passive (linear or nonlinear) operator.
For diagonal passive LTI uncertainty, multipliers are chosen to be real rational, diagonal,
bounded on the imaginary axis, and so that they satisfy
W (j!) = W (j!)� > 0; 8 ! 2 R:
The next subsection closely parallels the developments in [Feron, 1994]. We will focus on
the worst case H2 performance of the uncertain system (4). This analysis will be used to
formulate the controller synthesis problem in x3.4.
3.2 Worst Case H2 Performance
The robust performance analysis forms the foundation of the robust controller synthesis
presented in this paper. This subsection provides a brief overview of the performance analysis
involving generalized multipliers [Feron, 1994].
Consider the uncertain system (4). Let x0 be any initial condition with zero disturbance
input. Assume that the uncertain system (4) is stable. We are interested in computing
9
the worst case output energy for the system subject to real parametric uncertainty, i.e.,
Jx0 := max�2� kzk22. While this quantity is very di�cult to compute, we are interested in
an upper bound which can be calculated relatively easily. The following lemma gives an
upper bound on Jx0.
Lemma 3.1 ([Feron, 1994]) Consider the uncertain system (4). Let W be a family of multi-
pliers which have the form (5) and satisfy the assumptions of Theorem 3.1. Then the output
energy, Jx0, of the uncertain system (4) is bounded by
Jx0 � minW2W
maxp2L2
(kzk22 � 2Z1
0pT qW dt);
where z; p; and qW satisfy
_x = Ax + Bpp+ Bww + Buu
qW = Cqx+ Dqpp+ Dqww + Dquu
z = Czx + Dzpp+ Dzww + Dzuu
y = Cyx + Dypp+ Dyww + Dyuu
(7)
where xT = [ xT xTW ], xT0 = [ xT0 0 ], and
266666664
A Bp Bw Bu
Cq Dqp Dqw Dqu
Cz Dzp Dzw Dzu
Cy Dyp Dyw Dyu
377777775=
266666666664
A 0 Bp Bw Bu
BWCq AW BWDqp 0 0
DWCq CW DWDqp 0 0
Cz 0 0 0 Dzu
Cy 0 Dyp Dyw Dyu
377777777775:
Proof: See [Feron, 1994]. 2
We note that the term
maxp2L2
(�Z1
0pT qW dt) (8)
subject to (7) has the interpretation of the maximum extractable energy from p to qW in
state x0. Because of the relationship p = ��q, where � is strictly passive, W is passive, (8)
has a nonnegative value. Given W 2 W, computing the upper bound of the output energy
is equivalent to computing
maxp2L2
Z1
0
"x
p
#T "CTz Cz �CT
q
�Cq �(Dqp + DTqp)
# "x
p
#dt; (9)
where x and p satisfy (7). Therefore, computing the bound is simply a linear quadratic
optimal control problem which we will state in the following theorem. The solution could
10
be obtained by standard methods [Willems, 1971, Anderson and Moore, 1990].
Theorem 3.2 ([Feron, 1994]) Consider the uncertain system (4). Suppose there exists the
multiplier W of the form (5) satisfying the assumptions of Theorem 3.1. Then the upper
bound of the output energy is �nite and can be computed as the optimization problem of
minimizing x0P x0 over the variables P ; PWi; CWi
; DWisubject to
P = P T > 0;24 AT P + P A+ CT
z Cz P Bp � CTq
BTp P � Cq �(Dqp + DT
qp)
35 < 0;
9>>>=>>>; (10)
24 AT
WiPWi
+ PWiAWi
PWiBWi
� CTWi
BTWiPWi
� CWi�(DWi
+DTWi)
35 � 0;
PWi= P T
Wi> 0; 8 i 2 1; : : : ; np
9>>>=>>>; (11)
Proof: See [Feron, 1994]. 2
Note that the condition (10) implies the strictly positive real constraint of WGqp and the
condition (11) is equivalent to the positive real constraint of the multiplier W . Both (10)
and (11) are LMIs in the variables P ; PWi; CWi
; DWi. For notational convenience, we de�ne
PW := diagnpi=1(PWi
).
For the uncertain system (4), the H2 performance is derived from the total energy of the
performance outputs zi(t) subject to impulse disturbances �wi; i = 1; : : : ; nw. Consequently,
we de�ne the worst case H2 performance J
J := max�2�
nwXi=1
kzik22: (12)
While J is di�cult to compute, its upper bound can be easily computed. As discussed
in [Stoorvogel, 1993], the appropriate initial conditions xi(0); i = 1; : : : ; nw, used in the
computation of this upper bound, are equal to Bwei, where feignwi=1 is an orthonormal basis
of the input disturbance space Rnw . Applying Theorem 3.2 to the worst caseH2 performance
in (12), then J is bounded by
J � Tr BTwP Bw: (13)
In summary, to compute the minimum cost overbound of the output energy for the uncertain
system (4), we solve the following optimization problem which we refer to as the worst case
H2 performance analysis with multipliers.
minimize Tr BTwP Bw
subject to (10); (11):(14)
11
Remark 3.3 Although the analysis of the worst case H2 performance in this subsection is
parallel to the previous technique in [Feron, 1994], a major di�erence exists. In [Feron, 1994],
the analysis involves a noncausal multiplier, which makes the initial condition of the multi-
plier state not equal to zero but depends on the input p in (7). Hence, in order to compute
the upper bound of the output energy (9) it requires a multiple-step technique, which subse-
quently complicates the optimization procedure. In contrast, in this work we use the causal
multiplier of the form (5), which results in a zero initial condition of the multiplier state,
i.e., xW;0 = 0. As we will show in x3.4, the analysis formulation can be easily extended to
the synthesis problem, which results in a clean and concise presentation.
In the next subsection, we will present a systematic way of choosing a basis of the
multiplier by using the information of the open-loop transfer function (the linear time-
invariant part) of the uncertain system (4).
3.3 Multiplier Construction
In previous approaches [Safonov and Chiang, 1993, Balakrishnan, 1995, Feron, 1994], arbi-
trary basis functions for the multiplier are chosen without taking into account the knowl-
edge of the open-loop transfer function of the uncertain system (4). However, [Brockett and
Willems, 1965] provide an explicit expression for the multipliers that make WGqp strictly
passive for the case of a single uncertainty. Their approach uses the information of the
open-loop transfer function of the uncertain system to place poles of the multiplier, which is
referred to as a plant{dependent multiplier. In [Brockett and Willems, 1965], the multiplier
has the form
� + s�1Qi(s
2 + �2i )Qj(s2 + �2j )
;
where � > 0, �i and �j are the frequencies satisfying
arg[Gqp(j!)] = 0;
d arg[Gqp(j!)]
d !
(< 0; ! = �i;
> 0; ! = �j;
(15)
where arg[Gqp(j!)] is the phase of Gqp(j!). In this paper, we consider a family of multipliers
for which (AW , BW ) are �xed, and (CW , DW ) are free to vary, provided that the conditions
in Theorem 3.1 are satis�ed. The reason for this restriction will become clear when we im-
plement the robust performance analysis test. As suggested in [Brockett and Willems, 1965,
How and Haddad, 1994], we will place the eigenvalues of AW to the locations that have the
natural frequencies equal to �j satisfying (15). In order to make the multiplier contain no
poles on the imaginary axis, we slightly shift the eigenvalues of AW into the left half of the
s-plane by adding an arbitrary small damping ratio �W;j. Since the state-space realization of
the multiplier is arbitrary, we use a modal realization which automatically gives BW within
12
a constant. This idea of choosing (AW ; BW ) can be extended to the multiple uncertainty
case where (AWi; BWi
) are constructed by the information of the (i; i) element of Gqp.
For future reference, let Wf denote the class of multipliers having the form (5) and
(AW ; BW ) are computed using this selection algorithm.
In the next subsection, we will use the analysis tool presented in the previous section to
design an LTI controller such that the performance objective is satis�ed.
3.4 Multiplier Controller Synthesis
This paper shows, for the �rst time, the extension of the robustness tests with generalized
multipliers to the robust H2 synthesis. The robust H2 performance analysis can be used
to design robust controllers. The design objective is to �nd a strictly proper full order LTI
controller that minimizes the upper bound of the worst case H2 performance. The controller
is of the form_xc = Acxc +Bcy
u = Ccxc(16)
where xc : R+ ! Rn+nW is the controller state and (Ac; Bc; Cc) are constant matrices of
appropriate size. We proceed by �rst describing the closed loop system of the augmented
system (7) and the LTI controller (16) by
_~x = ~A~x+ ~Bpp+ ~Bww
qW = ~Cq~x+ ~Dqpp+ ~Dqww
z = ~Cz~x + ~Dzpp+ ~Dzww
(17)
where ~xT = [ xT xTc ], and
26664
~A ~Bp~Bw
~Cq~Dqp
~Dqw
~Cz~Dzp
~Dzw
37775 =
266666664
A BuCc Bp Bw
BcCy Ac +BcDyuCc BcDyp BcDyw
Cq DquCc Dqp 0
Cz DzuCc 0 0
377777775:
Then, it is straightforward to compute the upper bound of the worst case H2 performance for
the closed loop system (17). We note that the condition (10) is equivalent to ~P = ~P T > 0,
and "~AT ~P + ~P ~A + ~CT
z~Cz
~P ~Bp � ~CTq
~BTp~P � ~Cq �( ~Dqp + ~DT
qp)
#< 0; (18)
In summary, the design objective is to solve the following optimization problem:
minimize Tr ~BTw~P ~Bw
subject to (11); (18); ~P > 0:(19)
13
4 Design Procedure
We �rst note that (19) is a BMI problem, i.e., there are product terms involving the anal-
ysis parameters ( ~P , CW , and DW ) and compensator parameters (Ac, Bc, and Cc). The
formulation is quite complicated because it requires a simultaneous optimization of both the
multipliers and compensators. Observing the structure of the compensator parameters in
(19), the �rst step of the design procedure is to eliminate some controller parameters from
the problem formulation. We then solve for the remaining variables, and use these results to
construct the controllers. An iterative algorithm is required to calculate the compensators,
but this process capitalizes on the very e�cient design tools that are available for solving
LMI problems [Vandenberghe and Boyd, 1994, Wu and Boyd, 1996].
4.1 Controller Elimination
We �rst note that the controller matrix Ac only appears in the inequality (18). Thus it is
possible to reduce the number of variables in the optimization problem by eliminating Ac.
To proceed, we de�ne
~A0 :=
"A BuCc
BcCy BcDyuCc
#; ~J :=
"0
I
#:
Then ~A can be written as ~A = ~A0 + ~JAc~JT and we rewrite the inequality (18) as
~G+ V ATc U
T + UAcVT < 0; (20)
where ~G, U , and V are de�ned as
~G :=
24 ~AT
0~P + ~P ~A0 + ~CT
z~Cz
~P ~Bp � ~CTq
~BTp~P � ~Cq �( ~Dqp + ~DT
qp)
35 ;
U :=
"~P ~J
0
#; V :=
"~J
0
#:
Therefore, the orthogonal complements of U and V are
U? =
"~P�1 ~J? 0
0 I
#; V? =
"~J? 0
0 I
#:
Applying the Elimination Lemma, it follows that the inequality (20) holds if and only if
"~J? 0
0 I
#T~G
"~J? 0
0 I
#< 0;
"~P�1 ~J? 0
0 I
#T~G
"~P�1 ~J? 0
0 I
#< 0: (21)
14
To proceed, we partition ~P and its inverse ~Q as
~P =
"P M
MT R
#; ~Q = ~P�1 =
"Q N
NT S
#; (22)
where P and Q 2 R(n+nW )�(n+nW ). Then, after some algebra, it can be shown that the
inequalities (21) are equivalent to
24 PA+ ZCy + (PA+ ZCy)
T + CTz Cz PBp + ZDyp � CT
q
(PBp + ZDyp � CTq )
T �(Dqp + DTqp)
35 < 0;
26664AQ+ BuY + (AQ+ BuY )
T Bp �QT CTq (CzQ+ DzuY )
T
(Bp �QCTq )
T �(Dqp + DTqp) 0
CzQ + DzuY 0 �I
37775 < 0;
9>>>>>>>>>>>=>>>>>>>>>>>;
(23)
where Y and Z are de�ned as Y := CcNT ; Z := MBc. By the Completion Lemma, the
conditions ~P > 0; ~P ~Q = I with ~P given by (22) imply
"P I
I Q
#� 0: (24)
Restricting (24) to be positive de�nite, we are e�ectively searching for full-order controllers
(i.e., of order n+nW ) [El Ghaoui and Folcher, 1996]. We observe that the second inequality
in (23) is BMI, i.e., there are product terms involving Q and (CW ; DW ). This is a direct
consequence of optimizing both the compensator parameters (related to Q) and the analysis
multiplier (CW ; DW ) simultaneously. Note that if (CW ; DW ) are �xed, then the inequalities
(23) are LMIs in Q. Similarly, if Q is �xed, then the inequalities (23) are LMIs in (CW ; DW ).
Now we consider the objective function which is an upper bound of the worst case H2
performance, i.e., Tr ~BTw~P ~Bw. The trace objective is equivalent to
Tr
"Bw
Dyw
#T 24 P Z
ZT X
35"Bw
Dyw
#; (25)
and the existence of a symmetric matrix X such that
2664X ZT 0
Z P I
0 I Q
3775 > 0: (26)
Following [El Ghaoui and Folcher, 1996], we note that the inequality (26) implies ~P > 0.
In summary, after eliminating Ac from the formulation the optimization problem (19) is
15
equivalent tominimize (25)
subject to (11); (23); (26):(27)
4.2 Controller Reconstruction
Given that there exist P; Q; Y; Z; X; PW ; CW ; DW satisfying (27), we can construct a con-
troller using the following procedure. We �rst construct ~P such that the inequality (18)
holds. ~P is parameterized by (22), where M is an arbitrary invertible matrix. Because M
corresponds to a change of coordinates in the controller states xc, the choice of M has no
e�ect on the controller transfer function [El Ghaoui and Folcher, 1996]. After constructing~P , the set of input/output controller matrices (Bc; Cc) can be parameterized by Bc =M�1Z
and Cc = Y (I � PQ)�1M . With ~P ; CW ; DW ; Bc and Cc determined, it su�ces to �nd Ac
satisfying the inequality (20), which can then be formulated as an LMI problem in Ac.
4.3 Algorithm
It has already been shown that BMI problems are NP-hard, and it is thought to be rather
unlikely that there is a polynomial time algorithm for solving the general BMI problem
[Toker and �Ozbay, 1995]. Since there are product terms involving compensator parameters
and the multiplier parameters, our approach to solve the non-convex optimization problem
is based on an iterative procedure. First, we systematically select the multiplier dynamics
(AW ; BW ) by the method described in x3.3, i.e., W 2 Wf . The proposed algorithm, which
we call the V{K iteration, is basically an iteration between three di�erent LMI problems,
i.e., (19) with �xed compensator parameters, (27) with �xed multiplier parameters, and
(20) over Ac. The �rst LMI problem, considered as the V or analysis step, is to solve (19)
with �xed (Ac, Bc, and Cc) which yields multiplier parameters (CW and DW ). For the K or
synthesis step, the second and third LMI problems are solved. The solution parameters of
the second LMI problem, i.e., (27) with �xed multiplier parameters, implicitly contains the
input/output compensator matrices (Bc and Cc) as variables. After obtaining Bc and Cc, the
dynamics of the compensator Ac can be computed by solving the third LMI problem (20).
At this point, a robust compensator, which guarantees the robust stability and satis�es the
upper bound of the worst case H2 performance, is completely calculated. We then repeat
the procedure until the decrease in the upper bound of the worst case H2 performance is
su�ciently small. The solution algorithm to design a set of controllers for systems with real
parametric uncertainty satisfying j�ij < is brie y summarized in Table 1. As discussed in
x3, a bilinear sector transformation [Desoer and Vidyasagar, 1975] can be used to convert
this problem into a form in which the passivity theorem can be applied.
Remark 4.1 The procedure of alternating between the LMI problems is an iterative ap-
proach of solving a non-convex optimization problem. It is not guaranteed to converge in
16
general, but in our experience it does converge, although not necessarily to the global op-
timum. Note that each step of the iteration can be solved very e�ciently by a previously
developed semide�nite programming algorithm sp [Vandenberghe and Boyd, 1994] and very
easily coded using a user-friendly interface sdpsol [Wu and Boyd, 1996].
Remark 4.2 An important distinction between the V-K iteration and the D-K iteration of
the �=Km synthesis is that in our approach there is shared variable between each iteration:
speci�cally, ~P is the common variable between the V step and the K step, in which ~P appears
as P , Z, Q, Y , and X. However, for the D{K iteration the D step (the �=Km analysis) is
entirely separate from the K step (the H1 synthesis).
Table 1: Algorithm of Multiplier H2 controller synthesis.
1. Initialize the uncertainty to be zero (a nominal system) and design the con-troller via Linear Quadratic Gaussian (LQG) or any other robust controldesign technique.
2. Choose (AW , BW ) by a method such as one described in x3.3. Initialize (CW ,DW ) by solving (19) where (Ac, Bc, Cc) are �xed.
3. Repeatf [Outer Loop]
(a) Repeatf [Inner Loop]
i. Solve the optimization problem (27), i.e., solving for (P , Q, Y , Z,X, PW ) where (CW , DW ) are �xed. Then compute ~P , Bc, and Cc
using the Completion Lemma.
ii. Compute Ac by solving a feasibility LMI problem (20).
iii. Compute (CW , DW ) by solving (19) where (Ac, Bc, Cc) are �xed.
g [Inner Loop] Until stopping criterion satis�ed.
(b) Increase the uncertainty to the next desired size and initialize CW andDW by the most recent values.
g [Outer Loop] Until the desired robustness is achieved or the problem isinfeasible.
17
5 Numerical Example
In [Grocott et al., 1994], the authors compare several robust control design techniques using
benchmark problems based on a cantilevered Bernoulli Euler beam with unit length and mass
density, and sti�ness scaled so that the fundamental frequency is 1 rad/sec. The in�nite order
dynamics of the beam are truncated at four modes, where w1 = 1 rad/sec, w2 = 6:27 rad/sec,
w3 = 17:55 rad/sec, w4 = 34:39 rad/sec and damping � = 0:01. The disturbance input,
control input, sensor output and performance output are all collocated at the tip of the beam,
and the frequency of the third mode of the system is considered to be uncertain. With �5%
shifts in the modal frequency, there are substantial variations in the system gain and phase in
the 17� 25 rad/sec frequency range [Grocott et al., 1994, Banjerdpongchai and How, 1996].
The uncertainty � in this case is a constant real scalar that satis�es j�j < , where is
referred to as a guaranteed stability bound. As discussed in x3, we use a bilinear sector
transformation [Desoer and Vidyasagar, 1975] to convert this problem into a form in which
the passivity theorem can be applied.
The �rst step in the synthesis process was to design an LQG controller for the nominal
system, i.e., = 0, and then apply the technique described in x3.3 to specify multiplier
matrices AW and BW . Using the procedure in (15), the natural frequency is selected as
�W � 17 rad/sec. The damping ratio, �W , is arbitrarily set equal to 0:1. This selection
process explicitly shows the plant dependent nature of the stability multiplier. Several
controllers were designed using the LMI synthesis approach. Note that for this benchmark
problem, each iteration of the outer-loop in the algorithm in x4.3 required approximately 7
minutes to execute on a Sun-Sparc 20/60.
There are many interesting aspects to these robust controllers, but we restrict the discus-
sion to a comparison with Popov controllers [Banjerdpongchai and How, 1996]. Note that
both design techniques provide robust stability and performance guarantees for parameter
variations within the uncertainty region. However, there is an important distinction between
these two techniques. While Popov controllers are designed to capture the memoryless sector
bounded nonlinear uncertainty, multiplier controllers directly address real parametric uncer-
tainty (in this example). We will compare the H2 performance of both design techniques for
the same guaranteed stability bounds, i.e., comparing the Popov controller with the multi-
plier controller designed for the equal size of . This consistency is necessary to make a fair
comparison between two di�erent design techniques.
There are two distinct quantities that are often used to measure the conservatism of
di�erent design techniques. One measure is the increase of the H2 cost in the guaranteed ro-
bustness bounds from the nominalH2 cost (i.e., theH2 cost evaluated on the nominal system
with the LQG design). The smaller the increase of the H2 cost within the guaranteed region,
the less conservative the control design technique. A second measure of the conservatism
is the di�erence or gap between the guaranteed and achieved stability bounds. The ideal,
but not realizable, situation would be to have a non-conservative control design technique
18
that for any size of guaranteed bound yields compensators with a normalized H2 cost, i.e.,
the H2 cost normalized by the nominal H2 cost, equal to one within the guaranteed region.
Although wider achieved stability bounds indicate more robustness of the control designs,
the performance achieved outside the guaranteed regions is not directly addressed in the
design. Therefore, it could be bene�cial to sacri�ce the performance outside the guaranteed
regions to obtain a better performance inside.
Figure 2 depicts the results obtained for the normalizedH2 cost with the given percentage
change in the modal frequency. The robust H2 performance of three control design tech-
niques: LQG, Popov, and multiplier controller designs are compared. Table 2 summarizes
the key points of the robust performance from this plot: the percentage change of the H2
cost at the nominal system for Popov and multiplier controller designs compared with the
nominal H2 cost, and the lower (upper) achieved and guaranteed stability bounds. From the
plot, we note the following observations.
� Comparing the controllers designed using di�erent techniques for the same guaranteed
robustness bounds, the normalized H2 cost for the Popov controllers is signi�cantly
higher than that for the multiplier controllers in the guaranteed regions. This improve-
ment is clearly shown in the expanded plot in Figure 3. Since we would like to achieve
guaranteed robustness bounds with the minimum possible degradation in the nomi-
nal performance, these results indicate that the multiplier controller designs are less
conservative than the Popov ones designed for the same guaranteed stability bounds.
� For each design, the achieved stability bound is larger than the guaranteed bound but
generally these bounds track each other, i.e., a larger guarantee bound is accompanied
by a larger stability margin. Figure 2 and Table 2 show that, for the same guaranteed
bound, the achieved stability bounds for the multiplier controllers are smaller than
those achieved by the Popov controllers. Moreover, the di�erence or gap between the
guaranteed and achieved stability bounds of multiplier controllers is smaller than that
of Popov controllers. The narrower gap of the multiplier controller designs potentially
indicates that the control e�ort is concentrated on achieving improved performance of
the closed-loop system for uncertainty within the guaranteed region. This observation
is consistent with the overall design objective.
These two observations strongly support the claim that extending the Popov multipliers to
generalized multipliers reduces the conservatism in the robust performance analysis/synthesis
for a system with real parametric uncertainty. This is because the generalized multipliers can
better capture the uncertainty which is real and constant, while the Popov multiplier was
devised for a larger class of uncertainty, i.e., sector bounded nonlinearity, which considers real
parametric uncertainty as a special case. As a consequence, Popov controller designs yield
wider achieved stability bounds than multiplier controller synthesis when real parametric
uncertainty is under consideration.
19
�20 0 20 40
1
1:25
1:50
1:75
2LQG 2% 2% 4% 4%
Percentage Change in the uncertain Frequency
NormalizedH2Cost
LQG
PCS
MCS
Figure 2: Robust performance plots for LQG, PCS and MCS controllersdesigned with the symmetric robustness 2% and 4% bounds la-beled by the curves.
�10 �5 0 5 10
1
1:05
1:10
LQG
2%2%
4%4%
Percentage Change in the uncertain Frequency
NormalizedH2Cost
LQG
PCS
MCS
Figure 3: The expanded plot of Figure 2 showing the robust performanceabout the nominal frequency.
20
Table 2: Robust stability and performance for the closed-loop system. Forconsistency, it is necessary to make a fair comparison betweentwo di�erent design techniques for the same guaranteed stabilitybounds, i.e., comparing the Popov controller with the multipliercontroller designed for the equal size of uncertainty.
Type of % Change of Lower Stability Bound % Upper Stability Bound %Controller H2;nom Cost Achieved Guaranteed Guaranteed AchievedLQG 0 �5 0 0 7PCS2 2:36 �10 �2 2 16PCS4 4:22 �15 �4 4 34MCS2 1:28 �9 �2 2 13MCS4 3:28 �13 �4 4 23
We continue the comparison of the control techniques in terms of the pole and zero
location of various robust compensators. This includes the Popov and multiplier controllers
at several guaranteed stability bounds. Figure 4 shows that the multiplier controllers have
two distinct groups of poles and zeros in the uncertain region (the frequency range local
to the uncertain mode). Recall that the full-order multiplier controller has higher order
than the full-order Popov controller by two. The extra compensator poles are at a frequency
similar to that of the multiplier dynamics augmented to the system for the analysis test. The
�gure also shows that the corresponding Popov controllers have only one pole and one zero
in this range, and that these controller dynamics become heavily damped as the robustness
level is increased. On the other hand, one pole-zero pair of the multiplier controllers is more
lightly damped than the Popov controller design, while the second pole-zero pair, which was
lightly damped initially, becomes more heavily damped as the robustness level is increased.
From this plot, we also see that the di�erence between two control techniques outside of the
uncertain frequency region is quite small. Figure 5 shows the frequency response of LQG,
Popov and multiplier controllers for the uncertainty bound with = 0:04. These graphs
show that the di�erences in the pole-zero patterns in Figure 4 lead to subtle changes in
the frequency response of the compensators in the uncertain frequency region and at higher
frequency. Figure 5 also indicates that the phase of the compensators di�ers by as much
as 25� at approximately 15 rad/sec. These plots are interesting because they show how the
compensators have been changed, but of course, it is di�cult to directly identify how these
changes improve the robustness bounds of the closed-loop system.
To explore this last point further, a non-conservative real parametric robust analysis
must include both magnitude and phase information about the uncertainty �. This would
be evident in a Nyquist plot. In particular, the inverse of the minimum (maximum) real
axis intercept of the Nyquist plot can be used to determine the lower (upper) bound of the
21
real uncertainty that the closed-loop system can tolerate. We show the impact of various
compensators on a Nyquist plot of the transfer function from q to p (across the uncertainty
�) for the closed-loop system. To guarantee the uncertainty bound with = 0:04, the real
axis intercepts must lie between �25 and 25 on the Nyquist plot. Figure 6 shows that the
Nyquist plot of the multiplier controller synthesis almost always encircles the Nyquist plot
of the Popov controller synthesis. Furthermore, the real axis intercepts have been increased
towards their target of �25. Thus, as expected, the controller design with generalized
multipliers results in an improvement in the magnitude of the real axis intercepts. This
result demonstrates that the control e�ort of multiplier controllers is exerted more within
the guaranteed regions. This observation agrees with the claim that the conservatism of the
performance and achieved robustness bounds is reduced when the control synthesis with the
generalized multipliers is applied.
KW Poles
KW Zeros
KP Zeros
KP Poles
�1:5 �1 �0:5 0 0:50
10
20
30
40
Re
Im
�
�
�
�
�
Figure 4: Poles and zeros of the Popov controllers, KP and multiplier con-trollers, KW for the robustness bounds 2, 4, 6, 8 and 10%. Thearrows show the direction of change with increasing robustness.Large changes in the uncertain region are clearly evident.
22
LQG
MCS
PCS
PCS
MCS
LQG
10 20 30 40 50100
101
102
Magnitude
10 20 30 40 50�600
�400
�200
Frequency (rad/sec)
Angle(degree)
Figure 5: Frequency response of LQG, Popov and multiplier controllersrobusti�ed to frequency errors in the third mode. Robust con-trollers are designed for the uncertainty bound with = 0:04.Signi�cant changes to the response are apparent in the 10 � 20rad/sec range.
23
PCS
MCS
�8 �4 �0 4 8�8
�4
0
4
8
Re
Im
Figure 6: Comparison of the Nyquist plots of the closed-loop transfer func-tion from p to q (across the uncertainty �) for Popov and mul-tiplier controller design. Both controllers are designed for theuncertainty bound with = 0:04.
24
6 Conclusions
This paper presents an iterative technique for parametric robust H2 control design with gen-
eralized multipliers using LMI synthesis. These multipliers better capture information of the
uncertainty �, which helps reduce the conservatism in the associated robust performance
analysis tests for systems with real parametric uncertainty. This approach is limited because
not all of the multiplier parameters can be optimized during the current synthesis algorithm,
but a systematic procedure is provided for selecting the dynamics of multipliers using knowl-
edge of the uncertain systems. This multiplier selection algorithm is a signi�cant �rst step
towards a new, combined analysis and synthesis methodology which extends the prior work
on the robust stability and performance analysis. The design procedure is shown to be an
e�ective robust control design technique. In particular, we demonstrate that this new ap-
proach produces less conservative compensators than previous Popov controller techniques
for a Bernoulli Euler beam with an uncertain modal frequency. A signi�cant advantage of
LMI synthesis over previous procedure using gradient based optimization techniques is the
low overhead associated with developing the optimization conditions. This advantage greatly
simpli�es the numerical implementation for problems involving the simultaneous optimiza-
tion of multiplier and controller parameters.
Acknowledgments
This research was supported by Ananda Mahidol Foundation and in part by AFOSR under
grant F49620-95-1-0318. We would like to thank the anonymous reviewers for valuable
comments and suggestions.
References
[Anderson, 1972] Anderson, B. D. O. (1972). The Small-Gain Theorem, the Passivity andTheir Equivalence. J. Franklin Inst., 293(2):105{115.
[Anderson and Moore, 1990] Anderson, B. D. O. and Moore, J. B. (1990). Optimal Control:Linear Quadractic Methods. Prentice-Hall.
[Anderson and Vongpanitlerd, 1973] Anderson, B. D. O. and Vongpanitlerd, S. (1973). Net-work Analysis and Synthesis: a Modern Systems Theory Approach. Prentice-Hall.
[Balakrishnan, 1995] Balakrishnan, V. (1995). Linear Matrix Inequalities in Analysis withMultipliers. Syst. Control Letters, 25(4):265{272.
25
[Balakrishnan, 1997] Balakrishnan, V. (1997). Robust Performance Bounds Based on Lya-punov Funtions for Uncertain Systems. In Proc. Allerton Conf. on Circuit and SystemTheory. Preprint.
[Banjerdpongchai and How, 1996] Banjerdpongchai, D. and How, J. P. (1996). ParametricRobustH2 Control Design Using LMI Synthesis. In The 1996 AIAA Guidance, Navigation,and Control Conference, AIAA-96-3733.
[Boyd et al., 1994] Boyd, S., El Ghaoui, L., Feron, E., and Balakrishnan, V. (1994). Lin-ear Matrix Inequalities in System and Control Theory, volume 15 of Studies in AppliedMathematics. SIAM, Philadelphia, PA.
[Brockett and Willems, 1965] Brockett, R. W. and Willems, J. L. (1965). Frequency DomainStability Criteria | Part I & II. IEEE Trans. Aut. Control, AC-10(3, 4):255{261, 407{412.
[Desoer and Vidyasagar, 1975] Desoer, C. A. and Vidyasagar, M. (1975). Feedback Systems:Input-Output Properties. Academic Press, New York.
[Doyle, 1982] Doyle, J. (1982). Analysis of Feedback Systems with Structured Uncertainties.IEE Proc., 129-D(6):242{250.
[Doyle, 1983] Doyle, J. (1983). Synthesis of Robust Controllers and Filters. In Proc. IEEEConf. on Decision and Control, pages 109{114.
[El Ghaoui and Balakrishnan, 1994] El Ghaoui, L. and Balakrishnan, V. (1994). Synthesis ofFixed-structure Controllers via Numerical Optimization. In Proc. IEEE Conf. on Decisionand Control, pages 2678{2683.
[El Ghaoui and Folcher, 1996] El Ghaoui, L. and Folcher, J. P. (1996). Multiobjective Ro-bust Control of LTI Control Design for Systems with Unstructured Perturbations. Syst.Control Letters, 28(1):23{30.
[Fan et al., 1991] Fan, M. K. H., Tits, A. L., and Doyle, J. C. (1991). Robustness in thePresence of Mixed Parametric Uncertainty and Unmodeled Dynamics. IEEE Trans. Aut.Control, AC-36(1):25{38.
[Feron, 1994] Feron, E. (1994). Analysis of Robust H2 Performance with Multipliers. InProc. IEEE Conf. on Decision and Control, pages 2015{2020.
[Goh et al., 1994] Goh, K. C., Ly, J. H., Turand, L., and Safonov, M. G. (1994). �=km-Synthesis via Bilinear Matrix Inequalities. In Proc. IEEE Conf. on Decision and Control,pages 2032{2037.
[Grocott et al., 1994] Grocott, S. C. O., How, J. P., and Miller, D. W. (1994). Comparisonof Robust Control Techniques for Uncertain Structural Systems. In AIAA Guidance,Navigation, and Control Conference, pages 261{271.
26
[Haddad and Bernstein, 1991] Haddad, W. and Bernstein, D. (1991). Parameter DependentLyapunov Functions, Constant Real Parameter Uncertainty, and the Popov Criterion inRobust Analysis and Synthesis. In Proc. IEEE Conf. on Decision and Control, pages2274{2279, 2632{2633.
[Haddad et al., 1992] Haddad, W., How, J., Hall, S., and Bernstein, D. (1992). Extensions ofMixed-� Bounds to Monotonic and Odd Monotonic Nonlinearities Using Absolute StabilityTheory. In Proc. IEEE Conf. on Decision and Control.
[How, 1993] How, J. P. (1993). Robust Control Design with Real Parameter Uncertaintyusing Absolute Stability Theory. PhD thesis, Massachusetts Institute of Technology, Cam-bridge, MA 02139.
[How and Haddad, 1994] How, J. P. and Haddad, W. M. (1994). Robust Stability andPerformance Analysis using LC Multipliers and the Nyquist Criterion. In Proc. AmericanControl Conf., pages 1418{1419.
[How et al., 1994] How, J. P., Hall, S. R., and Haddad, W. M. (1994). Robust Controllers forthe Middeck Active Control Experiment Using Popov Controller Synthesis. IEEE Trans.Control Sys. Tech., 2(2):73{87.
[Packard et al., 1991] Packard, A., Zhou, K., Pandey, P., and Becker, G. (1991). A Collectionof Robust Control Problems Leading to LMI's. In Proc. IEEE Conf. on Decision andControl, pages 1245{1250.
[Safonov, 1982] Safonov, M. G. (1982). Stability Margins of Diagonally Perturbed Multi-variable Feedback Systems. IEE Proc., 129-D(6):251{256.
[Safonov, 1983] Safonov, M. G. (1983). L1 Optimal Sensitivity vs. Stability Margin. InProc. IEEE Conf. on Decision and Control.
[Safonov and Chiang, 1993] Safonov, M. G. and Chiang, R. Y. (1993). Real/Complex Km-Synthesis without Curve Fitting. In Leondes, C. T., editor, Control and Dynamic Systems,volume 56, pages 303{324. Academic Press, New York.
[Safonov et al., 1994] Safonov, M. G., Goh, K. C., and Ly, J. H. (1994). Control SystemSynthesis via Bilinear Matrix Inequalities. In Proc. American Control Conf., pages 45{49.
[Sandberg, 1964] Sandberg, I. W. (1964). On the L2-Boundedness of Solutions of Non-linearFunctional Equations. Bell Syst. Tech. J., 43:1581{1599.
[Sparks and Bernstein, 1995] Sparks, A. G. and Bernstein, D. S. (1995). Real StructuredSingular Value Synthesis Using the Scaled Popov Criterion. AIAA J. of Guidance, Control,and Dynamics, 18(6):1244{1252.
[Stoorvogel, 1993] Stoorvogel, A. A. (1993). The Robust H2 Control Problem: A WorstCase Design. IEEE Trans. Aut. Control, AC-38(9):1358{1370.
27
[Toker and �Ozbay, 1995] Toker, O. and �Ozbay, H. (1995). On the NP-Hardness of SolvingBilinear Matrix Inequalities and Simultaneous Stabilization with Static Output Feedback.In Proc. American Control Conf., pages 2525{2526.
[Vandenberghe and Boyd, 1994] Vandenberghe, L. and Boyd, S. (1994). sp: Software forSemide�nite Programming. User's Guide, Beta Version. K.U. Leuven and Stanford Uni-versity.
[Willems, 1971] Willems, J. C. (1971). Least Squares Stationary Optimal Control and theAlgebraic Riccati Equation. IEEE Trans. Aut. Control, AC-16(6):621{634.
[Wu and Boyd, 1996] Wu, S.-P. and Boyd, S. (1996). sdpsol: A Parser/Solver for Semide�-nite Programming and Determinant Maximization Problems with Matrix Structure. User'sGuide, Beta Version. Stanford University.
[Young, 1993] Young, P. M. (1993). Robustness with Parametric and Dynamic Uncertainty.PhD thesis, California Institute of Technology, Pasadena, CA 91125.
[Zames, 1966a] Zames, G. (1966a). On the Input-Output Stability of Time-Varying Non-linear Feedback Systems | Part I : Conditions Derived Using concepts of Loop Gain,Conicity, and Positivity. IEEE Trans. Aut. Control, AC-11:228{238.
[Zames, 1966b] Zames, G. (1966b). On the Input-Output Stability of Time-Varying Non-linear Feedback Systems | Part II: Conditions Involving Circles in the Frequency Planeand Sector Nonlinearities. IEEE Trans. Aut. Control, AC-11:465{476.
28